Less than 1.0 CPUs on a GPU task causes scheduler thrashing
16 Jan 2021 22:07:11 UTC
Topic 224520
(moderation:
Any time I get a GPU task wanting 0.5 CPUs or 0.9 CPUs, the scheduler tries to run one more task than there are CPU cores available.
Productivity collapses and task duration skyrockets, presumably due to the CPU cores having to swap tasks continually.
As long as all GPU tasks want 1 CPU, all is well.
The only solution I have found thus far is to deselect "Binary Radio Pulsar Search (Arecibo, GPU)" and "Gravitational Wave search O2 Multi-Directional GPU" so that only "Gamma-ray pulsar binary search #1 (GPU)" is being downloaded... but that solution lacks aesthetic appeal.
Worse, my newest computer, assigned the location "work", still downloads the Arecibo GPU tasks (wanting 0.5 CPUs) even though I've deselected them in the project preferences. I have to watch those like a hawk, and abort them when they show up. But this solution has even less aesthetic appeal.
Is there a way to force the scheduler to assign only one task per CPU core, regardless of what the task claims it needs?
Yes, you can do it with an
)
Yes, you can do it with an app_config.xml file placed in your einstein directory. See here: https://boinc.berkeley.edu/wiki/Client_configuration#Project-level_configuration
If you prefer to stick to
)
If you prefer to stick to preferences set at your account on the Einstein web site, then you could try setting:
Preferences|Computing|Processor Usage
Use at most: nn% of the processors
to just low enough a value of nn to get behavior more to your liking.
This affects the fetching assumptions, and affects the number and mix of tasks launched. It does nothing to influence how many CPUs the launched tasks actually run on.
I've found my way to that
)
I've found my way to that client-config file, and I see I can limit the CPU versions to 14 max-concurrent (out of 16 cores), as well as specifying 1 CPU per GPU task.
I'll have to do much the same for the oth3er computer as well, but that should hopefully get rid of the problem.
Thanks much!!
Reducing the number of
)
Reducing the number of processors available to BOINC doesn't offer the unused CPUs to the GPU tasks... it just forbids the use of some of the CPUs entirely.
Glenn Hawley, RASC Calgary
)
That turns out not to be the case. That number influences task fetch quantity, and task launches.
BOINC has no such behavior as forbidding use of CPUs. The assignment of tasks to CPUs is an OS function.
I put this point in to avoid that someone else coming on this thread might pick up wrong information. I recognize that you have found a mechanism you prefer.
archae86 wrote: BOINC has no
)
Any time I try , even putting % of CPUs at 99, only 15 tasks run, total.
Set at 90%, only 14 tasks run.